基于轻量化模型的人脸检测算法

Face Detection Algorithm Based on Lightweight Models

针对现有人脸检测算法的准确性主要依赖于大规模参数的问题,提出了一种基于单阶段人脸检测算法Retinaface网络的人脸检测算法。该算法使用GhostNet作为Retinaface网络的特征提取网络,利用线性计算节约计算资源,以期提高计算效率;使用加权双向特征金字塔网络加强特征图之间的信息交互,增强模型的特征融合能力。在Widerface数据集上对原模型与改进后的模型进行测试,结果表明,与其他模型相比,改进后模型的精度较原模型提升最高7.35%。

A face detection algorithm based on the single-stage Retinaface network is proposed to address the accuracy issue in face detection algorithms that rely on large-scale parameters.The GhostNet serves as the feature extraction network of the Retinaface network,and the linear computation conserves computational resources and enhances computational efficiency.The information exchanges between feature maps are fortified and the model's feature fusion capabilities are elevated with weighted bidirectional feature pyramid network.The original and optimized models are tested on the Widerface dataset,and experimental results show that the accuracy of optimized model improved by up to 7.35%over the original model.